## Machine Learning at Cambridge Computer Lab --- ### New faculty * Neil Lawrence * Carl Henrik Ek * Ferenc Huszár --- ### ML@CL  --- # Introductions --- ### ML opportunities at Computer Lab * PhD in Computer Science (3+ years) * MPhil in Advanced CS (1 year, research-heavy) * [AI for Environmental Risk](https://ai4er-cdt.esc.cam.ac.uk/) CDT (1+3 years) * [Accelerate Programme for Scientific Discovery](https://www.cst.cam.ac.uk/news/new-programme-accelerate-ai-research-capability-cambridge) * research groups * ML@CL * Artificial Intelligence Group * Natural Language Processing --- ### ML opportunities at Cambridge * [MLG at Engineering Department](http://mlg.eng.cam.ac.uk/) * [Machine Intelligence Lab](http://mi.eng.cam.ac.uk/) at Engineering * computer vision, speech, robotics * [MPhil in Machine Learning and Machine Intelligence](https://www.postgraduate.study.cam.ac.uk/courses/directory/egegmpmsl) * Microsoft Research, Amazon, Samsung, Huawei, nvidia, etc --- ## Our research interests --- ## Ferenc: ### deep learning methodology 1. ### representation learning 1. ### optimization and generalization 1. ### probabilistic foundations 1. ### causal inference --- ## representation learning * Goals and principles: what makes a representation good? * * self-supervised learning * transfer learning * meta-learning * few-shot learning --- ## optimization and generalization * Why deep networks generalize? * implicit regularization by stochastic gradient descent * natural gradient descent * neural tangent kernels * infinite width neural networks * differential privacy --- ## probabilistic foundations * What's the best form of inference in deep networks? * generalized Bayesian methods * cold/tempered posteriors * exchangeability in deep learning * probabilistic interpretation of non-probabilistic algorithms * continual learning --- ## causal inference * How can deep learning learn causal models? * deep learning for causal inference * causal theory for robust deep learning * invariant risk minimization * deep instrumental variable --- ## what I'm not working on ### (with exceptions) * regret bounds, learning therory (COLT, ALT, etc) * specific applications * computer vision, natural language processing * reinforcement learning/robotics --- ## What you need for a PhD #### Maths background * linear algebra * probability theory * calculus and vector calculus * statistics * graphs --- ## What you need for a PhD #### Programming * python * numerical computing (numpy, scipy) * tensor libraries (pytorch, jax, tensorflow) --- ## What you need for a PhD #### Intuitions about ML concepts * KL-divergence, entropy, variational bounds * probability distributions and their properties * conditional independence, statistical independence * Bayesian inference * optimization: convex, non-convex, constrained * matrix decompositions --- ## What you need for a PhD #### Research interests * not too specific * not too broad * you have opinions about your area of ML * you can identify specific research questions --- ## What you need for a PhD #### Past experience * experience in working on a research project * a paper is nice but not a requirement * samples of writing or code are useful * ideally: explain what you learned, and how it shapes your interests --- ## Funding and scholarships * first an academic offer, then funding offer * scholarship should cover: * tuition fees * maintenance grant ("salary") * several scholarship options * can be very complicated * DeepMind Scholarships * 3 MPhil students * 2 PhDs * women and underrepresented groups --- ## Cambridge application process * apply via website * there's an application fee * reach out to supervisor informally first * recommendation letters * cover letter/research statements * language test * transcripts --- ## Thanks! [inference.vc/phd](https://www.inference.vc/information-for-prospective-phd-students/)