FARRUKH BARLAS

Prompt engineering remains essential for extracting optimal performance from transformer-based LLMs without parameter updates. Core techniques include Chain-of-Thought (CoT) prompting, which decomposes complex problems into intermediate reasoning steps, and Tree-of-Thoughts (ToT), which implements tree search over multiple reasoning trajectories.   Key Uses: In software engineering, few-shot and role-based prompts enhance code generation accuracy in models like GPT-4 or Claude. In healthcare, structured prompts improve named entity recognition (NER) from clinical notes.  Hacks: Use delimiters (XML-style tags) to prevent prompt injection. Use delimiters (XML-style tags) to prevent prompt injection. Implement prompt chaining for multi-step workflows. Apply Automatic Prompt Engineer (APE) where a meta-LLM iteratively optimizes prompts using evolutionary search. Adjust temperature (0.0–1.0) and top-p nucleus sampling for creative vs deterministic outputs. References: Wei, J., et a...

Adversarial examples exploit the sensitivity of deep neural networks (DNNs) to imperceptible perturbations. The Fast Gradient Sign Method (FGSM) computes adversarial examples via   x′=x+ϵ⋅sign(∇x​J(θ,x,y))  Key Uses: Testing robustness of computer vision models in autonomous vehicles (e.g., attacking YOLO detectors) and malware classification systems.  Hacks & Defences:  Adversarial training augments the training dataset with perturbed samples generated by Projected Gradient Descent (PGD). Use feature squeezing and randomization at inference time. Black-box attack hack: Train a substitute model to approximate the target’s decision boundary. Physical-world hack: Generate robust adversarial patches that survive real-world lighting variations.  References: Goodfellow, I.J., et al. (2014). Explaining and Harnessing Adversarial Examples. arXiv:1412.6572. Madry, A., et al. (2018). Towards Deep Learning Models Resistant to Adversarial Attacks. ICLR

Federated Learning (FL) enables collaborative model training across decentralized devices while keeping raw data local. The central server aggregates model updates (typically weights or gradients) using FedAvg algorithm.  Key Uses:  Healthcare (training diagnostic models across hospitals) and mobile keyboard prediction without uploading user text.  Hacks:  Implement differential privacy by adding Gaussian noise to gradients. Use secure multi-party computation (SMPC) or homomorphic encryption for model aggregation. Apply model compression (sparsification) before transmission to reduce bandwidth. Hierarchical FL for cross-silo and cross-device scenarios.  References: McMahan, B., et al. (2017). Communication-Efficient Learning of Deep Networks from Decentralized Data. AISTATS.

Reinforcement Learning (RL) trains agents via reward signals. Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC) are popular for continuous control tasks.  Key Uses: Robotic manipulation, drone navigation, and quadruped locomotion.  Hacks:  Use domain randomization during simulation training for better sim-to-real transfer. Apply Hindsight Experience Replay (HER) for sparse reward problems. Curriculum learning to gradually increase task difficulty. Combine with imitation learning (behaviour cloning) for sample efficiency.  References: Schulman, J., et al. (2017). Proximal Policy Optimization Algorithms. arXiv:1707.06347.

TinyML brings machine learning to microcontrollers with severe memory (<256KB RAM) and compute constraints. Post-training quantization converts FP32 weights to INT8. Key Uses: Wake-word detection, predictive maintenance sensors, and always-on vision on embedded devices.  Hacks:  Dynamic range quantization and float16/INT8 mixed precision. Pruning (magnitude-based weight removal) followed by fine-tuning. Knowledge distillation from large teacher to tiny student model. Use TensorFlow Lite Micro or Edge Impulse for optimized deployment.

Contrastive Language-Image Pretraining (CLIP) learns joint embeddings by contrasting positive and negative image-text pairs using a dual-encoder architecture.  Key Uses: Zero-shot image classification, content moderation, and visual search.  Hacks:  Use CoOp (Context Optimization) to learn continuous prompt vectors. Fine-tune with LoRA on custom datasets for domain adaptation. Combine with visual question answering (VQA) pipelines. Leverage CLIP embeddings for unsupervised clustering of visual data.  References:  Radford, A., et al. (2021). Learning Transferable Visual Models From Natural Language Supervision. ICML.

SHAP (SHapley Additive exPlanations) assigns importance values to features based on cooperative game theory. LIME approximates the model locally with an interpretable surrogate.  Key Uses: Financial risk models, medical diagnosis systems, and regulatory compliance.  Hacks:  Use Kernel SHAP for model-agnostic explanations on tabular data. Visualize attention maps in transformers as an additional interpretability layer. Combine LIME with counterfactual explanations ("what if" analysis). Implement partial dependence plots alongside SHAP for global insights.  References: Lundberg, S.M., & Lee, S.I. (2017). A Unified Approach to Interpreting Model Predictions. NeurIPS.