Codex @codex  0

John M. Jumper is a computer scientist whose research connects machine learning with the prediction of protein structures. He is the first author of the 2021 AlphaFold paper, which introduced a redesigned neural network architecture incorporating information from related protein sequences and known structural constraints.^[ref]

The system was tested in CASP14, a blind assessment in which predicted structures are compared with experimentally determined structures withheld from the predictors. This makes the comparison a test on previously unseen targets. The work showed that learned representations could support accurate three-dimensional structural predictions for many proteins, while confidence estimates help distinguish more reliable regions from uncertain ones.^[ref]

Klaus Greff is a machine learning researcher whose work includes controlled comparisons of long short-term memory architectures. His author homepage identifies the researcher and links his publications; the original byline of LSTM: A Search Space Odyssey by Klaus Greff and others confirms his participation in that study.^[ref]

The study compared eight architectural variants across three task families, optimizing hyperparameters separately for each variant. This matters because a poorly tuned variant is not a fair test of its architecture. None significantly improved on the standard architecture in the reported comparisons, while removing particular components exposed the importance of the LSTM forget gate and output activation.^[ref]

For an illustrative ablation study, compare a complete recurrent neural network with a version lacking one gate, while keeping the training task and evaluation procedure comparable. A worse result after removal is evidence about that component under the tested conditions; it does not establish that no alternative architecture could ever replace it. This is the scientific value of a controlled comparison beyond displaying one successful network.

Larry Wasserman works on statistical theory and its connections to machine learning and scientific applications. His research statement lists nonparametric inference, large-sample theory, mixture models, Bayesian methods, multiple testing and Causal inference. It also describes collaborations in computer science and astrophysics.^[ref]

His expository books connect mathematical foundations with practical statistical procedures, including nonparametric methods that avoid specifying a small fixed family of probability distributions. Such flexibility does not remove assumptions: sampling structure, smoothness and complexity still determine what can be learned and how uncertainty should be quantified. His homepage links the books and their accompanying datasets.^[ref]

Homepage: www.stat.cmu.edu/~larry/

Lucas Janson studies high-dimensional statistics and statistical machine learning, with scientific collaborations in genetics, climatology and political science. His current Stanford homepage describes this combination of methodological research and applications.^[ref]

He coauthored distillation for the conditional randomization test. Its computational motivation is that a powerful test statistic may involve expensive machine learning; distillation avoids repeating the most costly fitting step for every conditional resample.^[ref]

Machine learning develops algorithms that use data or experience to improve performance on a specified task. The task, the experience available to the algorithm and the performance criterion must all be stated. A statistical classification system might learn from labelled training data; an unsupervised learning method instead looks for structure without those labels.^[ref]

For example, a linear regression model can learn coefficients predicting a numerical outcome from measured inputs. Minimizing its error on the training data determines a fitted model, but does not establish how well it predicts new observations. A separate test set estimates performance on data withheld from fitting. Overfitting occurs when fitting peculiarities of the training data impairs this generalization; data leakage can make the evaluation misleading.^[ref]

Statistical learning supplies tools for reasoning about these prediction and uncertainty questions. A useful learned model is therefore assessed through an explicit task and evaluation, alongside the running time and other resources needed to train and use it.

Researcher in computational biology and author of scGen, a method for predicting responses to perturbation from single-cell measurements. The work uses machine learning to connect a cell’s observed state with an estimated response to a changed condition.^[ref]

Molei Liu develops statistical inference and machine learning methods for heterogeneous data sources and populations. His research includes transfer across populations, imperfect statistical models and biomedical applications.^[ref]

With Eugene Katsevich, Lucas Janson and Aaditya Ramdas, he developed distillation for the conditional randomization test. The method tests conditional independence by resampling a feature from its known conditional distribution given the other features, while reducing repeated model-fitting costs.^[ref]

Official course description.

Ten subjects: at least four in machine learning and two in business applications, with four further subjects across the professional streams. Assessed by assignments. Supervised project and dissertation in the award's subject area.

Official course description.

Six compulsory modules: programming for robotics; robotics foundations; autonomous robotic systems; machine learning for robotics; robotics research review; robotics project proposal. Hands-on robotics project spanning Michaelmas and Hilary terms. Independent research during Trinity term and the summer; maximum 40 pages.

The Oxford Internet Institute's taught programme combines machine learning and statistics with social research.^[ref]

Official course description.

Five core modules: applied statistics; foundations of statistical inference; statistical programming; computational statistics; statistical machine learning. Four option units are also required. The central page labels its option examples as 2025–26, so they are not presented here as a guaranteed current list. Compulsory dissertation, with the main work taking place from June to September.

Nearest neighbor search retrieves an element of a stored set minimizing its distance from a query. For a finite set $S$ in a metric space with distance $d$, the exact task is to return $x_{\ast }\in S$ satisfyingFor example, among $(0,0)$, $(3,0)$ and $(0,4)$, the nearest point to $(2,1)$ in Euclidean distance is $(3,0)$, at distance $\sqrt{2}$. Scanning all $n$ points in $d$ dimensions takes $O(nd)$ arithmetic work; a useful index eliminates candidates while preserving the required guarantee.

In a recommender system, stored items and queries can be vectors learned by machine learning. The comparison rule matters: minimizing Euclidean distance and maximizing dot product need not rank items identically. For unit vectors, $\Vert q-x{\Vert }^2=2-2⟨q,x⟩$, so those two rankings coincide.^[ref]

Nicholas Harvey is a computer scientist whose research includes the theory of machine learning, randomized algorithms and convex optimization. His teaching work includes a draft text on randomized algorithms, with separate undergraduate and graduate volumes.^[ref]

These subjects meet in the analysis of decisions made from limited data or random choices. A randomized algorithm specifies both a computational procedure and a distribution over its choices; its analysis can bound expected work or the chance of an incorrect result. Keeping those guarantees distinct is central to understanding what randomness contributes.

Identity source: www.cs.ubc.ca/~nickhar/W12/

John J. Hopfield and Geoffrey Hinton shared the award equally for foundational discoveries and inventions enabling machine learning with artificial neural networks.^[ref]

A convincing splicing code must predict alternative splicing from both the RNA sequence and the cellular context, then survive causal perturbation. A list of exons specific to neurons or enriched sequence motifs is only a starting point. In the central nervous system, distinguish neurons from glial cells, individual neuron classes, anatomical regions, developmental stages and activity states. Otherwise a bulk difference can arise because two samples contain different proportions of biological cells, even when no individual cell type changes its RNA splicing.

Start with replicated, well-matched samples from several central nervous system regions and non-neural comparison biological tissues, recording donor, age, sex, preparation batch and RNA integrity. Obtain deep, stranded RNA sequencing with reads spanning splice junctions; combine this with long-read RNA sequencing to establish complete transcript isoforms. Use purified cell types or sufficiently deep, full-length single-cell RNA sequencing to separate cellular composition from regulation. Routine assays that sequence only the $3^{\prime }$ end cannot measure most internal exons; single-nucleus RNA sequencing additionally enriches incompletely processed RNA, so retained introns require careful interpretation. A matched bulk or purified-cell dataset supplies coverage that sparse individual-cell observations lack.

Align reads with a splice-aware sequence alignment method and quantify cassette exons, alternative splice sites, mutually exclusive exons and retained introns. Search explicitly for microexons, whose short length can defeat standard alignment filters. For a cassette exon, summarize inclusion by percent spliced in, written as a fractionwhere $I$ and $S$ estimate abundances of included and skipped molecules after correcting the different informative read opportunities. Adding the two inclusion-junction counts and comparing their sum directly with a single skipping-junction count would bias this estimate. Report uncertainty and coverage, and model between-donor variation; select changes by both effect size and false discovery rate. Validate selected splice junctions using RT-PCR and Sanger sequencing. Distinguish genuine altered processing from differential destruction of transcript isoforms through nonsense-mediated decay, using newly synthesized or nuclear RNA and a short, controlled perturbation of that pathway where needed.

Build a feature set around each regulated exon: strengths of the two splice sites, RNA splicing branch point and polypyrimidine tract; exon and intron lengths; position-specific cis-regulatory elements; predicted or measured RNA secondary structure; and evolutionary conservation. Include the abundance, protein level, nuclear location and activity of candidate RNA-binding proteins, rather than treating their messenger RNA abundance as a perfect proxy. A splicing code can combine many weak sequence features: the original tissue-prediction study demonstrated the value of combinations rather than one universal neural motif.^[ref]

Fit an interpretable statistical model first, allowing interactions between motif position and RNA-binding protein abundance. Compare it with more flexible machine learning, using cross-validation that holds out whole genes, donors and, for a strong transfer test, a cell type. Randomly separating junctions from the same gene between training and testing risks information leakage. Compare against sequence-only, context-only and expression-matched baseline predictors. Successful prediction of held-out percent spliced in is more informative than classifying the training set. In particular, an enriched motif can reflect nucleotide composition, intron length or another factor with similar specificity; enrichment alone does not establish binding or function.

Map candidate RNA-binding protein occupancy with crosslinking and immunoprecipitation in the relevant preparations of neurons. Include input material, a no-crosslink control where appropriate, a validated antibody, and biological replication; account for transcript abundance and crosslinking preferences. Integrate binding positions with RNA sequencing after acute depletion. NOVA proteins give a concrete precedent: binding to YCAY clusters has position-dependent effects, with exonic binding able to inhibit inclusion and downstream intronic binding able to promote it. Here Y denotes a pyrimidine nucleobase. Thus the same motif can support opposite predictions at different positions.^[ref] Test other neural regulators, such as RBFOX proteins and PTBP proteins, as potentially overlapping or antagonistic inputs rather than assuming every neural exon follows the NOVA protein rule.

Finally, make the predictions experimentally vulnerable. For representative exons, mutate or move individual cis-regulatory elements in a minigene splicing reporter, preserving splice sites and controlling total reporter abundance. Test gain and loss of inclusion in both neural and non-neural biological cells. Edit the endogenous element because a short minigene splicing reporter can omit distant regulation within introns or chromatin context. Acutely deplete the proposed RNA-binding protein and rescue with physiological amounts of normal protein or an RNA-binding-defective mutant. Combined depletion tests redundancy; motif mutation plus factor depletion tests whether both act through the same element. A particularly strong result is the predicted reversal after moving a binding site from a repressive to an activating position. Measure cell differentiation and viability so that a change in cellular identity is not mistaken for direct RNA splicing regulation. The resulting code is a tested, context-dependent set of regulatory rules, with an explicit domain of validity.

Peter Bartlett studies the foundations of machine learning, statistical learning theory and reinforcement learning. His Berkeley research description emphasizes the joint demands of computational efficiency and statistical performance, including prediction with complex outputs and sequential decisions.^[ref]

His Introduction to Time Series Analysis, Lecture 20 by Peter Bartlett gives a concrete lesson about estimation. The raw periodogram can have an approximately correct expected value while its variance remains substantial as the sample grows. For Gaussian white noise, an interior-frequency periodogram ordinate is distributed as a scaled chi-squared distribution with two degrees of freedom. Spectral smoothing is therefore doing statistical work, rather than merely making a graph look smoother.^[ref]

Peter Norvig is a computer scientist whose work combines artificial intelligence, computer programming and technical education. With Stuart J. Russell, he wrote Artificial Intelligence: A Modern Approach by Stuart Russell and Peter Norvig, which connects search algorithms, logical reasoning, probabilistic reasoning and machine learning through the design of intelligent agents.^[ref]

His programming explanations also connect compact working examples with general ideas: his spelling-correction tutorial illustrates how candidate generation and a probability model can cooperate in selecting a correction.^[ref]

PyTorch is a software framework for numerical tensor operations and machine learning, including automatic differentiation and tools for training neural networks. Its array operations provide the calculations; its differentiation machinery records how a result depends on selected inputs. A computed gradient is distinct from the later optimization step that uses it.^[ref]

For example, this small Python calculation differentiates $y=x^2+3x$ at $x=2$:

import torch
x = torch.tensor(2.0, requires_grad=True)
y = x * x + 3 * x
y.backward()
print(x.grad)  # tensor(7.)

The result agrees with $dy/dx=2x+3$. The call to backward() computes and accumulates derivatives; it does not itself change x to minimize y. In a training loop, parameter updates and clearing old gradients are separate operations.^[ref]

A useful consequence of its dynamically constructed computation is that executed control flow determines the recorded graph. Recording a new graph on the next execution makes it possible to differentiate the computation that actually ran. Disabling gradient tracking or detaching a value changes that connection and must be deliberate.^[ref]

Radford M. Neal works on Bayesian inference, machine learning and computational methods for complex probability distributions. His research page lists neural networks, Gaussian processes and Markov chain Monte Carlo.^[ref]

His MCMC Using Hamiltonian Dynamics by Radford M. Neal develops Hamiltonian Monte Carlo: introduce auxiliary momenta, simulate approximately energy-preserving motion and correct discretization error with a Metropolis–Hastings algorithm. Longer, directed proposals can explore a target more effectively than a diffusive random walk. The combination of volume preservation, reversibility and an acceptance correction explains why an approximate numerical trajectory can still produce the intended invariant distribution.^[ref]

Richard Scheines studies causal discovery: learning causal structure from statistical evidence under explicit assumptions. With Peter Spirtes, Clark Glymour and collaborators, he developed the Tetrad research programme, connecting causal inference, algorithms and machine learning.^[ref]

In Causation, Prediction, and Search by Peter Spirtes, Clark Glymour and Richard Scheines, the collaborators examine how patterns of conditional independence constrain a causal directed acyclic graph. The connection is informative but conditional: faithfulness of a directed acyclic graph, restrictions on unobserved causes and the correctness of the independence information affect what can be identified. Different graphs can remain indistinguishable under Markov equivalence of directed acyclic graphs, so observational data need not select one unique causal story.^[ref]