Lecture 10 · AI4ST

Incremental & Semi-Supervised Learning

Models that keep learning after deployment: catastrophic forgetting and its mitigations, semi-supervised strategies, active learning, and activity discovery. The exam's favorite deck — three of the 21 simulation MCQs come from these slides.

⏱ ~55 min 📚 6 sections ✅ 12 MCQ + 2 open-ended ⭐ Includes sim Q4 · Q11 · Q12 — heaviest-hit lecture 🧠 Interactive mind map →
1

Static vs Adaptive Models

Every model so far was "static": trained once, with a predetermined set of classes, unable to evolve with new classes or new sensor patterns. But after deployment, temporal data streams continuously evolve.

The four changes in data streams

Sibling set — drill the differences
These four definitions are textbook exam material: learn to tell them apart by their one distinguishing phrase.
ChangeSlide definition
Concept driftThe target subject may perform activities with some differences (e.g., young vs elderly)
Concept evolutionNew patterns emerge from data: new activities or new ways of performing an activity
Concept forgettingTo improve both accuracy and efficiency, it is crucial to forget patterns if not representative of the target user
Outliers / anomaliesNew data may present sporadic samples particularly distant from learned patterns — distinguishing outliers from anomalies is crucial

Why not re-training from scratch?

Periodically re-training as labeled data arrives is not scalable: it requires continuously storing labeled data (original training set + new data) and periodic re-training that is computationally costly and cannot adapt quickly to stream changes.

Incremental (online) learning
Periodically updating the classifier with new mini-batches of training data — no need for re-training! The model continuously learns new patterns, new classes, new domains. (Static ML: learn once, deploy once. Adaptive ML: learn continually, deploy continually.)

Why does this matter in Ambient Intelligence?

  • Adapting to specific users (concept drift): changes in health conditions, routines/habits, living environments and sensing setups (feature evolution).
  • Labeled data scarcity: the HAR model starts from a small unrepresentative dataset; new labeled and unlabeled data arrive incrementally; it's challenging to foresee all activities (→ activity discovery, §6).
  • The "lifelong learning" loop: change detection → change annotation → change adaptation.
2

Catastrophic Forgetting

The major problem
Neural networks suffer from catastrophic forgetting of old concepts as new ones are learned: learning on new data would likely override the weights learned in the past, so the network may significantly degrade the recognition rate on concepts learned in the past.

The stability-plasticity dilemma

Stability

  • The capability of the model of retaining old knowledge

Plasticity

  • The capability of the model of learning and adapting on new knowledge

The dilemma: how to achieve stability without rigidity, and plasticity without chaos? Goal: a good trade-off between continuously learning and maintaining old information — a challenging open problem in AI.

Continual Learning (CL) studies learning from an infinite stream of data — to adapt models to new domains, learn new classes, personalize models — gradually extending acquired knowledge and using it for future learning. CL approaches mitigate catastrophic forgetting by balancing plasticity and stability.

Desirable ingredients for continuous learning: constant memory, undefined set of classes, incremental learning, few-shot learning, retaining and improving acquired knowledge, learning also from unlabeled data, selective (graceful) forgetting.

3

Mitigations: Replay & Regularization

Replay methods

Alleviate forgetting using "old" labeled data: maintain a limited subset of stored samples (replay memory) representing old knowledge, provided to the classifier along with new data at each incremental update. Pros: low-cost, easy to implement. Cons: risk of overfitting over stored samples.

Sibling set — the 4 replay variants (sim-Q14-style question material!)
The professor builds whole MCQs out of lists like this one — every option a sibling. Know each variant by its keyword.
VariantHow the buffer works
Random replayThe buffer is populated by randomly sampling previous labeled data points; when learning on new data, buffer data are also used
Exemplars replayThe buffer is populated with exemplars: representative labeled samples (e.g., samples close to the feature mean of each class)
Constrained-based replayThe last K data samples fill the buffer; learning on new data is constrained by a loss focused on old data: it can decrease but it can not increase
Generative replayWhen storing samples is unfeasible: the output of previous versions of the model with synthetic input approximates "old" knowledge (e.g., a GAN generating high-quality data). Cons: complexity significantly increased

Problems of replay-based methods: requires extra storage, the model may overfit on the buffer, buffer data may not be representative.

Regularization-based methods

Avoid storing raw inputs (prioritizing privacy) and alleviate memory requirements: extra regularization terms in the loss function consolidate previous knowledge. Two families:

Data-focused — LwF

  • Learning without Forgetting leverages knowledge distillation (Lecture 7b!)
  • The previous model (teacher) is maintained: it generates pseudo-labels on incoming data related to old tasks
  • A specific loss considers pseudo-labels + classification on new tasks: the student outputs a similar probability distribution on old knowledge and, at the same time, learns the new tasks

Prior-focused — EWC, VCL

  • Estimate an importance value for each weight (i.i.d. assumption)
  • While learning on new data, a regularization term penalizes significant changes to important weights
  • Important weights stay similar after training → knowledge maintained. Famous: Elastic Weight Consolidation, Variational Continual Learning
The exam's favorite discrimination
Sim Q11's four options are exactly: replay buffer (correct), LwF distillation, weight importance, and storing all data (the non-scalable strawman). If you can label each mechanism with its family in one read, the question is free points.
ThinkA hospital deploys a HAR model and forbids storing any raw patient sensor data on the device. Which catastrophic-forgetting mitigations remain available, and which are excluded?
Answer: Excluded: random, exemplars and constrained-based replay — all store raw old samples in a buffer. Available: the regularization-based family, which exists precisely to "avoid storing raw inputs, prioritizing privacy": LwF (only the previous model is kept, generating pseudo-labels on incoming data) and prior-focused methods like EWC (only per-weight importance values are kept). Generative replay is a borderline case: it stores no real samples (a GAN synthesizes them), at the price of significantly increased complexity.
4

Semi-Supervised Learning

Definition — sim Q12 lives here
In real life it's unlikely to obtain new labels from the sensor stream (labeled data scarcity). Semi-supervised methods: a small training set initializes the model before deployment; once deployed, the pre-trained model provides labels to unlabeled data points; the obtained labels (pseudo-labels) are used to update the model. Goal: reaching results close to fully supervised approaches. Note: pseudo-labels UPDATE the model — re-training from scratch is NOT "always necessary" (that absolutized claim is exactly Q12's FALSE option).
  • Personalization: a model trained collaboratively by volunteers → the target user downloads it locally → semi-supervised learning adapts it to the user.
  • Semi-supervised can be applied to static models too (when re-training from scratch is possible); here we focus on the incremental setting — more realistic in AmI.
  • The three main strategies: self-training, co-training, active learning (§5).

Self-training

Train on small labeled set Predict pseudo-labels + confidence on unlabeled data Keep only reliable (high-confidence) predictions Update the model
  • Label propagation variant: data points close in space likely share the class. Build a graph G=(V,E) — nodes are (un)labeled data points, edges weighted by similarity; labels spread from labeled to unlabeled points based on a threshold on similarity, repeated until convergence.
  • Label propagation cons: it must store unlabeled data points — impossible to store them all incrementally, so it runs periodically over a limited buffer of historical data.
  • Self-training limits: errors in label spreading may be amplified with time; hard to spread labels to data "different" from the training data.

Co-training

  • An ensemble of classifiers pre-trained on the available labeled data, each trained on a different view of the dataset (different feature vectors, different portions…).
  • On each unlabeled point every model predicts; each model uses the most confident predictions from the other models as additional pseudo-labeled samples to improve on its own view.
  • Cons: less effective when the views are not complementary (designing views is not trivial!); like self-training, errors in one model may be propagated to the others.
5

Active Learning

Definition — sim Q4 lives here
A semi-supervised approach widely used in AmI that requires interaction with the user: when the model is "uncertain" about a specific output (e.g., analyzing the probability distribution during classification), it triggers a query to the user, who provides the ground truth; the new labeled point updates the model. Such methods should minimize the number of triggered queries. The goal is querying only the most informative data points — the ones where the classifier is the most uncertain.

Uncertainty measures

Sibling set — the 3 measures
Three formulas, three rationales — perfect material for a sibling-harvest MCQ.
MeasureFormulaRationale
Least confidenceφLC(x) = 1 − Pθ(y*|x)The simplest: query when the confidence on the most likely class is below a threshold
Smallest marginφM(x) = Pθ(y₁*|x) − Pθ(y₂*|x)Difference between the two most confident predictions — how much more confident was the model on the top class vs the runner-up?
Entropy−Σᵢ P(yᵢ|x;θ) log P(yᵢ|x;θ)Considers the whole distribution: low entropy = one dominant class; high entropy = several activities with similar probabilities → query when above a threshold

Worked example (3 activities): [0.8, 0.15, 0.05] → entropy ≈ 0.61, A1 clearly dominates, no query. [0.5, 0.4, 0.1] → entropy ≈ 0.94, the classifier isn't sure about A1 → a query may refine the model.

Adaptive threshold (VAR-UNCERTAINTY)

  • Choosing a fixed threshold is hard in real scenarios → adjust it dynamically: when the system is uncertain, the threshold is adjusted to trigger more queries (the uncertainty region is decreased, θ←θ(1−s)); when certain, the region is widened (θ←θ(1+s)) and fewer queries fire.
  • It uses least confidence, but adapts to any measure; entropy is often more robust with several classes (e.g., HAR).

Pool-based vs Stream-based

Pool-based (general ML)

  • A large pool of unlabeled data; in an offline phase the system chooses the points maximizing uncertainty measures

Stream-based (AmI!)

  • A human can't label a sensor sample long after it's collected → uncertainty is evaluated on each new sensor sample, and the query fires as soon as the model is uncertain

Acceptability — the real-world wall

  • Most literature assumes users are constantly available — not realistic. Notifications at inappropriate moments cause stress and frustration, increase the feedback error rate, and create disaffection with the system.
  • Fixes: trigger queries considering the user's context (current activity, location, social context, recent queries); interruptibility models (person + context descriptors → interruptibility level); incentivize users (gamification, economic rewards, promise of a better model). Still open, poorly investigated problems!
Limits of the three strategies — summary slide, verbatim
Self-training: errors amplified with time; hard to spread labels to "different" data. Co-training: more robust, but similar problems + challenging to design effective views. Active learning: the most effective one, but requires user interaction which may not always be possible.
ThinkTwo predictions over 4 activities: P₁ = [0.45, 0.44, 0.06, 0.05] and P₂ = [0.45, 0.20, 0.20, 0.15]. Least confidence treats them identically — which measure separates them, and which one should trigger a query first?
Answer: Both have the same top confidence (0.45), so φLC = 0.55 for each — least confidence can't tell them apart. The smallest margin does: P₁ has margin 0.45−0.44 = 0.01 (a near tie between the top two!), P₂ has 0.45−0.20 = 0.25. By margin, P₁ is far more uncertain and should be queried first. Entropy instead weighs the whole distribution and would rank P₂ as more uncertain (probability mass spread over more classes). That's exactly why the choice of uncertainty measure matters — and why the slides note entropy is often more robust when there are several classes.
6

Activity Discovery & Evaluation

Activity discovery

  • A fixed set of activities can't be foreseen in advance — different users perform different activities. Semi-supervised learning may enable activity discovery!
  • Idea: pre-train on a dataset with fixed activities, then apply activity discovery methods to automatically find new patterns from unsupervised data points. New patterns may indicate new activities or different ways of performing previously considered activities (= concept evolution from §1!).
Cluster unlabeled stream data Discover reliable patterns Pre-trained model annotates clusters (self-training) Model uncertain on a cluster → may be a new activity!
  • The annotation of new clusters can be based on active learning.
  • Incremental activity discovery is significantly harder: besides the pre-trained model, it must memorize the original data distribution to avoid catastrophic forgetting and find sensor patterns deviating from the original distribution (new concepts). A replay approach: predict & discover outliers → in-memory buffer → semi-supervised clustering + label propagation → extend the classifier, create class prototypes, update the replay memory.

Evaluating incremental classifiers

  • Since they continuously learn, evaluate the evolution of the recognition rate. For each stream sample: classify with the current model, apply the semi-supervised method (e.g., check if an AL query is needed), store prediction + ground truth + pseudo-labels, memorize triggered queries, update the model when needed (with CL strategies).
  • Visualize with a sliding window of the predictions: per window compute the metrics of interest (e.g., F1 score, number of queries). The window size sets the evaluation periodicity — it should be high enough that captured samples are representative.

Final Quiz — Exam Style

12 MCQs + 2 open-ended. Three are the actual simulation questions (Q11, Q12, Q4) verbatim; the rest are built with the professor's own recipe — sibling distractors, one-word swaps, a 2×2 grid, and one FALSE question hiding an absolutizer.

1
Which of the following describes "concept evolution" in data streams?
AThe target subject may perform activities with some differences (e.g., young vs elderly)
BNew patterns emerge from data: new activities or new ways of performing an activity
CIt is crucial to forget patterns if they are not representative of the target user
DNew data may present sporadic samples particularly distant from learned patterns
B — verbatim slide definition. The distractors are its siblings from the same two slides: A = concept drift, C = concept forgetting, D = outliers/anomalies. This four-way list is exactly the kind the exam harvests.
2
In the stability-plasticity dilemma, what is "plasticity"?
AThe capability of the model of retaining old knowledge
BThe capability of the model of compressing its weights
CThe capability of the model of forgetting selected patterns
DThe capability of the model of learning and adapting on new knowledge
D — verbatim. A is the one-word-swap trap: that's stability. The dilemma: "stability without rigidity, plasticity without chaos" — continual learning balances the two to mitigate catastrophic forgetting.
3
Simulation exam · Q11
How does the replay mechanism mitigate catastrophic forgetting in continual incremental learning?
ABy maintaining a small representative buffer of labeled data points about past information. When updating the model with new data, the data in the buffer are also used.
BBy distilling knowledge from the previous model encoding past information to a new model trained on new data.
CBy learning the importance of each model's weight. When updating the model on new data, a loss penalizes the changes of the most important weights.
DBy continuously storing all the data points used to train the model. When updating the model with new data, the old data are also used.
A — the slide: "maintaining a limited subset of stored samples (replay memory)… provided to the classifier along with new data at each incremental update". The distractors are real mechanisms from neighboring slides: B = LwF knowledge distillation (data-focused regularization), C = prior-focused regularization (EWC), D = the non-scalable re-training strawman ("continuously store labeled data") — a buffer is limited, not "all the data points".
4
Which of the following describes "exemplars replay"?
AThe buffer is populated by randomly sampling previous labeled data points.
BThe last K data samples fill the buffer, and a loss on old data can decrease but can not increase.
CThe buffer is populated with representative labeled data samples (e.g., samples close to the feature mean of each class).
DThe output of previous versions of the model with synthetic input is used to approximate old knowledge.
C — verbatim, parenthetical included (the exam keeps the slides' "(e.g., …)" examples!). Siblings from the same slide: A = random replay, B = constrained-based replay, D = generative replay. Same construction as the real Q14 on quantization: all four options are the four bullets of one slide.
5
Which of these sentences on replay-based methods is FALSE?
AThe replay memory is provided to the classifier along with new data at each incremental update
BReplay-based methods do not require any extra storage on the device
CThe model may overfit on the data stored in the buffer
DThe data in the buffer may not be representative of the old knowledge
B is FALSE — the slide lists "requires extra storage" as the first problem of replay-based methods; "do not require any" is the slide fact pushed to an absolute. A, C and D are verbatim true statements. Exam rule: in a FALSE question, hunt the option containing only / always / never / do not / any.
6
How does Learning without Forgetting (LwF) mitigate catastrophic forgetting?
ABy keeping a limited buffer of representative old samples used along with the new data
BBy penalizing, through a regularization term, significant changes to the most important weights
CBy training a GAN that generates synthetic data approximating the old knowledge
DBy keeping the previous model as a teacher that generates pseudo-labels on incoming data, with a loss making the new model match them while learning the new tasks
D — LwF is the data-focused regularization approach leveraging knowledge distillation: the previous model (teacher) generates pseudo-labels on incoming data related to old tasks; the loss ensures the student outputs a similar probability distribution on old knowledge while learning the new tasks. A = replay, B = prior-focused (EWC), C = generative replay — this question is sim Q11 with the roles rotated: drill the set in every direction.
7
What is the idea of prior-focused regularization methods (e.g., Elastic Weight Consolidation)?
AEstimate an importance value for each weight, and penalize significant changes to important weights while learning on new data
BEstimate an importance value for each class, and penalize significant changes to important classes while learning on new data
CMaintain the previous model to generate pseudo-labels for the data related to the old tasks
DFreeze all the layers of the network except the final classification head during the updates
A — verbatim: importance value per weight (i.i.d. assumption) + a regularization term penalizing significant changes to important weights, so important weights stay similar and knowledge is maintained (EWC, VCL). B is the one-word swap (weight→class); C is LwF, the other regularization family; D is plausible but appears nowhere in the deck — a fabricated mutation, exactly the kind of "modern-ML-sounding" distractor the sim uses.
8
Simulation exam · Q12
Which one of the following sentences on semi-supervised learning is FALSE?
AIt requires to pre-train a model with a limited amount of labeled data
BIt is a technique to mitigate labeled data scarcity
COnce pseudo-labels are obtained, it is always necessary to re-train the whole model from scratch
DIts goal is to spread pseudo-labels to unlabeled data points
C is FALSE — the slide says pseudo-labels "are used to update the model"; re-training the whole model from scratch is exactly what incremental learning avoids. The absolutizer "always" is the tell — the same trick as sim Q6 ("only BLE") and Q20 ("do not depend"). A, B, D are true slide statements.
9
In co-training, how does each classifier of the ensemble improve itself?
AIt uses its own most confident predictions as additional pseudo-labeled samples on its view
BIt uses the most confident predictions from the other models as additional pseudo-labeled samples on its view
CIt is re-trained on the majority vote computed across all the models of the ensemble
DIt exchanges its network weights with the other models after each mini-batch update
B — each classifier is trained on a different view of the dataset (different feature vectors, portions…), and uses the most confident predictions from the other models as extra pseudo-labels for its own view. A is the own↔other swap — that's self-training. Cons: views must be complementary (non-trivial design) and errors still propagate between models.
10
Simulation exam · Q4
How does Active Learning select the samples to be labeled?
Aby randomly choosing data points from the stream
Bby choosing the data points where the model is more uncertain
Cby choosing the data points that are different to the ones in the training set, according to a similarity function and a threshold
Dby choosing the data points that are better captured by common-sense knowledge
B — the slide: trigger a query only for the most informative data points, "the ones where the classifier is the most uncertain". Distractor C is a sibling from the same deck: similarity + threshold is label propagation; D imports knowledge-based vocabulary from Lecture 9. Note the correct answer is the shortest option — option length is never a tell.
11
Using the entropy uncertainty measure, when is an active learning query triggered?
AWhen the entropy is lower than a threshold, indicating one class with high probability
BWhen the entropy is higher than a threshold, indicating one class with high probability
CWhen the entropy is lower than a threshold, indicating several activities with similar probabilities
DWhen the entropy is higher than a threshold, indicating several activities with similar probabilities
D — a 2×2 grid: decide each axis separately. Axis 1: the user is queried when the entropy is higher than a threshold. Axis 2: high entropy means more activities have similar probability values (low entropy = one dominant class). Example: [0.8, 0.15, 0.05] → ≈0.61, no query; [0.5, 0.4, 0.1] → ≈0.94, query. Same structure as sim Q13 and Q18 — halve the problem twice.
12
Why is stream-based active learning preferred over pool-based in Ambient Intelligence?
ABecause the pool of unlabeled sensor data would be too large to store on a server
BBecause uncertainty measures cannot be computed offline on sensor data pools
CBecause a human can hardly label a sensor data sample long after it is collected
DBecause stream-based queries are guaranteed to be fewer than pool-based ones
C — pool-based AL (typical in Computer Vision) selects, in an offline phase, the points of a large unlabeled pool maximizing uncertainty. In AmI "it is likely impossible that a human can label a sensor data sample long after it is collected" — so uncertainty is evaluated on each new sample and the query fires as soon as the model is uncertain. (And remember the acceptability wall: ill-timed queries cause stress, higher feedback error rate, disaffection.)
13
Open-ended · Continual learning
a) Define catastrophic forgetting and the stability-plasticity dilemma. Why is re-training from scratch not a viable alternative?
b) Compare replay-based and regularization-based mitigations of catastrophic forgetting, describing their variants and trade-offs.
Model answer

a) Catastrophic forgetting: neural networks forget old concepts as new ones are learned — learning on new data would likely override the weights learned in the past, significantly degrading the recognition rate on past concepts. The stability-plasticity dilemma: stability is the capability of retaining old knowledge, plasticity the capability of learning and adapting on new knowledge; the dilemma is achieving "stability without rigidity and plasticity without chaos" — a trade-off between continuously learning and maintaining old information (an open problem in AI, which Continual Learning addresses by balancing the two). Re-training from scratch is in principle possible but not scalable in large real-world scenarios: it requires continuously storing labeled data (original training set + new data) and periodically re-training the model, which is computationally costly and cannot adapt quickly to changes in the stream. Incremental learning instead updates the classifier with mini-batches of new data.

b) Replay-based: maintain a limited subset of stored samples (replay memory) representing old knowledge, provided to the classifier along with new data at each incremental update. Variants: random replay (buffer filled by randomly sampling past labeled points), exemplars replay (representative samples, e.g., close to the feature mean of each class), constrained-based replay (last K samples + a loss on old data that can decrease but cannot increase), generative replay (when storing samples is unfeasible: previous model versions with synthetic input — e.g., a GAN — approximate old knowledge; complexity significantly increased). Pros: low-cost and easy to implement; cons: extra storage, the model may overfit on the buffer, buffer data may not be representative. Regularization-based: avoid storing raw inputs (prioritizing privacy) and alleviate memory requirements via extra regularization terms in the loss. Data-focused (LwF): knowledge distillation — the previous model (teacher) generates pseudo-labels on incoming data related to old tasks; a specific loss makes the new model (student) output a similar probability distribution on old knowledge while learning the new tasks. Prior-focused (EWC, VCL): estimate an importance value for each weight; a regularization term penalizes significant changes to important weights, so knowledge is maintained. Trade-off: replay is simple but stores raw data; regularization is privacy-friendly and memory-light but relies on approximations of past knowledge.

14
Open-ended · Semi-supervised learning & active learning
a) What is semi-supervised learning and why is it needed in AmI? Describe self-training (with label propagation) and co-training, with their limits.
b) Describe active learning: the uncertainty measures, stream-based vs pool-based operation, and the acceptability problem in real-world scenarios.
Model answer

a) In real-life streams it is unlikely to obtain new labels (labeled data scarcity). Semi-supervised learning: a small labeled training set initializes the model before deployment; the pre-trained model then provides labels (pseudo-labels) to unlabeled data points, and the pseudo-labels are used to update the model — aiming at results close to fully supervised approaches (it also enables personalization: a collaboratively trained model is downloaded by the target user and adapted locally). Self-training: the model predicts pseudo-labels with a confidence level on unlabeled data; only reliable (high-confidence) predictions are kept and used to update the model. A common implementation is label propagation: data points close in space likely share the label — build a graph whose nodes are labeled+unlabeled points and whose edges are weighted by similarity; labels spread from labeled to unlabeled nodes based on a similarity threshold, repeated until convergence. Limits: errors in label spreading are amplified with time; it is hard to spread labels to data "different" from the training data; storing unlabeled points is impossible incrementally (periodic runs on a limited buffer). Co-training: an ensemble of classifiers, each trained on a different view of the dataset (feature subsets, portions); each model uses the most confident predictions of the other models as additional pseudo-labels on its own view. Limits: less effective with non-complementary views (designing views is not trivial) and errors propagate between models.

b) Active learning — the most effective semi-supervised strategy — requires interaction with the user: when the model is uncertain about an output, it triggers a query; the user provides the ground truth and the new labeled point updates the model; the number of queries must be minimized by targeting only the most informative points (where the classifier is most uncertain). Uncertainty measures: least confidence φLC=1−P(y*|x) (query when top-class confidence is below threshold), smallest margin φM=P(y₁*|x)−P(y₂*|x) (gap between the two most confident classes), entropy −ΣP log P (whole distribution; high entropy = several classes with similar probabilities → query above a threshold; entropy is often more robust with many classes). Thresholds can be adaptive (VAR-UNCERTAINTY: shrink the uncertainty region when uncertain to trigger more queries, widen it when certain). Pool-based vs stream-based: general ML picks the most uncertain points offline from a large unlabeled pool; in AmI a human can't label sensor data long after collection, so uncertainty is evaluated on each new sample and queries fire as soon as the model is uncertain (stream-based). Acceptability: assuming users are constantly available is not realistic — ill-timed notifications cause stress and frustration, increase the feedback error rate and cause disaffection. Mitigations: trigger queries based on the user's context (activity, location, social context, recent queries), use interruptibility models, and incentivize labeling (gamification, rewards, promise of a better model) — still open research problems.

0/12
MCQ score 0/12