Description
There’s a peculiar tension at the heart of any book that promises to demystify something by showing you the math: the math might turn out to be the easy part. Anil Ananthaswamy’s Why Machines Learn spends roughly four hundred pages lovingly tracing the mathematical lineage from Rosenblatt’s perceptron to modern deep neural networks, only to arrive at an admission that the mathematics he’s just taught you can no longer explain why the machines actually work. It’s an odd structural gambit, though I’m not sure he intended it as one.
The premise is that a lay reader—armed with some high school algebra and a willingness to endure notation—can follow the conceptual arc of machine learning from vectors and dot products through gradient descent, Bayesian classification, support vector machines, convolutional neural networks, and on into the dizzying territory of large language models. Ananthaswamy, a veteran science journalist, proceeds gently, almost pedagogically, tethering each mathematical concept to a historical portrait. You get Frank Rosenblatt and his 1958 press conference, John Hopfield borrowing energy landscapes from spin glasses, Yann LeCun adapting Hubel and Wiesel’s neuroscience findings, the stubborn Geoffrey Hinton organizing bootleg workshops when no one would let him present on neural networks at official conferences. The human stories are, in fact, the book’s strongest current. When Hopfield describes how his training in physics led him to see neural computation as a dynamical system converging to stable states, or when Ilya Sutskever recalls his bewilderment at the simplicity of deep learning’s foundations—”How can it be that it’s so simple…so simple that you can explain it to high school students without too much effort?”—Ananthaswamy captures something real about the culture of this research: the vertigo of watching centuries-old math suddenly animate silicon.
The mathematical exposition itself is careful, occasionally too careful. Having spent years building statistical models for clinical research before drifting toward writing, I found the early chapters on vectors and calculus earnest but slow, covering ground that felt over-narrated for anyone who retained their undergraduate linear algebra and under-explained for anyone who hadn’t. The Monty Hall problem gets a multi-page treatment. Penguin species classification serves as the running illustration for probability and Bayesian reasoning—charming enough, though Ananthaswamy tends to repeat conceptual explanations he’s already given, as if afraid the reader has blinked. He acknowledges this impulse openly, comparing his own learning process to the iterative training of a neural network, each pass through the material deepening his grasp. Fair enough. But a book is not a network, and prose that cycles over the same ground without adding velocity can lose a reader who trusted it the first time.
Where the book sharpens is in its later chapters, particularly the treatment of “double descent” and the phenomenon called grokking. Mikhail Belkin and his colleagues showed that massively overparameterized neural networks—models with far more tunable knobs than data points—don’t fail the way classical learning theory predicts. Standard wisdom said such models should memorize training data and collapse on anything new. Instead, past a certain threshold, their test performance improves again. Belkin calls this uncharted mathematical territory “terra incognita,” and Ananthaswamy does well to convey both the excitement and the disorientation. “We [had] convinced ourselves that [ML] was fine by selectively closing our eyes on things that didn’t fit the mold,” Belkin tells him. That sentence could serve as the book’s quiet thesis, and it complicates the subtitle’s promise of “elegant math” in ways Ananthaswamy seems to sense but doesn’t quite confront. Elegance, after all, implies a complete picture. What he actually describes is a discipline whose practitioners have spent decades building a cathedral of theory only to discover its tallest spires rest on foundations they can no longer see.
This is where I find myself disagreeing with the book’s framing, or at least wanting more from it. Ananthaswamy writes in the prologue that “the denouement of this book leads us to a place that some might find disconcerting, though others will find it exhilarating,” and compares the moment to the early twentieth century’s break with classical physics. The comparison is evocative, but quantum mechanics arrived with new mathematics—Hilbert spaces, operators, wave functions—that replaced what had broken. Machine learning’s “terra incognita” has no equivalent replacement yet, only empirical observations outrunning theory. To spend hundreds of pages teaching readers the old theory and then tell them it no longer applies raises a question the book skirts: What exactly has the reader been equipped to evaluate? The treatment of bias, fairness, and the social harms of deployed systems—Google’s gorilla-tagging debacle, ProPublica’s recidivism investigation, Amazon’s sexist hiring algorithm—arrives in the epilogue with the compressed urgency of a writer who has run out of runway. These issues, which arguably need mathematics of their own (causal inference, counterfactual reasoning, distributional robustness), get a few pages against the twelve devoted to Lagrange multipliers.
None of this erases what the book does well. As an intellectual history of machine learning, read alongside the proofs and equations that drove it, Why Machines Learn has no real peer at this level of accessibility. The chapter on Hopfield networks alone—linking Ising models, associative memory, and energy minimization into a single narrative—is worth the cover price for anyone curious about how physics colonized computer science. And Ananthaswamy is honest, which counts for more than polish. He doesn’t pretend LLMs are solved or that pattern matching is reasoning. He quotes Emily Bender’s “stochastic parrots” label alongside researchers who see glimmers of genuine comprehension, and he leaves the question unresolved. That restraint is harder than it looks. Still, I keep returning to Belkin’s image of a map with blank spaces where the theory should be—cartography of ignorance, drawn in elegant lines that terminate at the edge of what anyone currently knows how to say.
If you’d like to read the full book in EPUB or MOBI format, feel free to send me an email—I’d be happy to share a free copy with you. Please reach me at: thenovaleaf@gmail.com
Reviews
There are no reviews yet.