Research Frontiers & Practice

Q1. (4 marks)

• Reproducibility = the ability of an independent party to obtain the same (or statistically consistent) results using the paper's described methods, given the same data/setup. (1)
• Three artifacts (1 mark each, any three): source code, datasets / data-splits, trained weights / checkpoints, hyperparameter configs & random seeds, environment specification (dependencies). (3)

Q2. (4 marks)

• InfoNCE/NT-Xent measures agreement between two augmented views of the same image, pulling positives together and pushing negatives apart in embedding space. (1)
• Loss for positive pair $(i,j)$: (3) $\ell_{i,j} = -\log \frac{\exp(\text{sim}(z_i,z_j)/\tau)}{\sum_{k=1}^{2N}\mathbb{1}_{[k\neq i]}\exp(\text{sim}(z_i,z_k)/\tau)}$ (1 for numerator = positive term, 1 for denominator = sum over all others, 1 for $\tau$ scaling.)

Q3. (3 marks)

• Positives per anchor: 1 (its other augmented view). (1.5)
• Negatives per anchor: $2N-2$ (all remaining views in the batch). (1.5)

Q4. (4 marks)

• CLIP trains on (image, text) pairs; a batch of $N$ pairs gives an $N\times N$ similarity matrix. The objective maximizes cosine similarity of the $N$ correct (matched) pairs and minimizes it for the $N^2-N$ mismatched pairs (symmetric cross-entropy over rows and columns). (2)
• Two encoders: an image encoder (e.g., ResNet/ViT) and a text encoder (Transformer). (2)

Q5. (4 marks)

• Catastrophic forgetting = when a model trained sequentially on new tasks loses performance on previously learned tasks because new-task gradients overwrite important weights. (2)
• Two families (1 each): regularization-based (e.g., EWC), replay/rehearsal-based (memory buffer), architecture/parameter-isolation (progressive nets). (2)

Q6. (5 marks)

• Federated learning = training a shared model across many decentralized clients that keep their raw data local; only model updates/weights are shared with a server. (2)
• FedAvg aggregation: (3) $w \leftarrow \sum_{k=1}^{K} \frac{n_k}{n}\, w_k$ (weighted average by sample count; 1 for weighting, 1 for sum over clients, 1 for correct normalization by $n$.)

Q7. (4 marks)

• A world model is a learned internal model that predicts how the environment evolves given the agent's actions, enabling planning/imagination without acting in the real world. (2)
• Two components: a dynamics/transition model ($s_{t+1}\sim p(\cdot|s_t,a_t)$) and an observation/representation (encoder) model (mapping observations to latent states); a reward model is also acceptable. (2)

Q8. (3 marks)

• Neuro-symbolic AI combines neural networks (learning from data) with symbolic reasoning (logic/rules/knowledge) in one system. (2)
• One advantage (any): better interpretability, data efficiency, generalization/compositionality, or ability to enforce hard constraints. (1)

Q9. (3 marks, 1 each)

• Test set may be too small / not representative or leaked from training (data leakage).
• Single metric hides class imbalance — accuracy misleading, no precision/recall/per-class breakdown.
• No confidence interval / statistical significance / multiple seeds — could be run-to-run luck; also possible train–test overlap or overfitting to the benchmark.

Q10. (6 marks)

• Three portfolio items (1 each, any three): reproduced/from-scratch model implementations, well-documented GitHub projects, blog posts / paper write-ups, Kaggle or benchmark results, published papers, a reproduced-paper repo. (3)
• Contribution methods (3 marks — any one clearly stated): start with "good first issue" / documentation fixes / writing tests / fixing small bugs, then submit a clear PR following the project's contribution guidelines. (3)

[
  {"claim": "SimCLR: with N=8 images, an anchor has 1 positive and 2N-2 = 14 negatives", "code": "N=8; positives=1; negatives=2*N-2; result = (positives==1 and negatives==14)"},
  {"claim": "FedAvg weights sum to 1: for n_k=[10,30,60], normalized weights sum to 1", "code": "nk=[10,30,60]; n=sum(nk); weights=[x/n for x in nk]; result = (sum(weights)==1)"},
  {"claim": "FedAvg of scalar client weights [2,4] with counts [10,30] gives 3.5", "code": "wk=[2,4]; nk=[10,30]; n=sum(nk); w=sum(Rational(nk[i],n)*wk[i] for i in range(2)); result = (w==Rational(7,2))"},
  {"claim": "CLIP batch of N=5 pairs has N=5 positives and N**2-N=20 negative pairs", "code": "N=5; pos=N; neg=N**2-N; result = (pos==5 and neg==20)"}
]