Convolutional neural networks quietly run your photo app, your car's cameras, and half the AI headlines you scroll past, but almost nobody outside a lab actually understands how they work. So we went through our entire library of podcast and lecture summaries and pulled out the episodes that explain CNNs best, straight from the people who built them.
This list mixes big-name interviews with full technical lectures. You get Yann LeCun, one of the actual inventors of the convolutional net, explaining why he wrote his own Lisp compiler at Bell Labs to make it work. You get Andrej Karpathy tracing the field from 1960s cat experiments to the 2012 breakthrough that changed everything. And you get Lex Fridman's entire MIT course on the subject, lecture by lecture, if you want the deep-dive version. Pick based on how far into the weeds you want to go.
Yann LeCun: Deep Learning, ConvNets, and Self-Supervised Learning | Lex Fridman Podcast #36
LeCun co-invented the convolutional neural network, so this is as close to a primary source as podcasting gets. He reveals that neural nets fell out of favor around 1995 partly because AT&T's lawyers blocked open-source release of the code, and that he and his team ended up writing their own Lisp interpreter and compiler just to build the original LeNet character-recognition system at Bell Labs. He also argues that the most surprising fact in deep learning is that gigantic over-parameterized networks work at all, breaking every rule from the pre-deep-learning textbooks. Anyone who wants the origin story and the philosophy behind CNNs, not just the math, should start here.
Read the full episode notesDeep Learning for Computer Vision (Andrej Karpathy, OpenAI)
Karpathy built Stanford's CS231N course on convolutional networks, and this lecture is the clearest historical thread on the list: Hubel and Wiesel's 1960s cat experiments, Fukushima's neurocognitron, LeCun's 1990s LeNet, and the 2012 AlexNet result that turned computer vision upside down. He points out that in roughly 20 years, the only two algorithmic advances over LeNet were dropout and ReLU, both one-line changes that just set values to zero, which is a genuinely humbling detail about how the field actually progresses. Listen to this one if you want the full lineage of CNN architecture explained by someone who taught it to a generation of researchers.
Read the full episode notesMIT 6.S094: Computer Vision
This lecture builds CNN intuition from scratch, starting with a trivial pixel-difference classifier that hits 35-38% on CIFAR-10 (versus a 10% random baseline) and climbing all the way to ResNet, which surpassed human-level ImageNet accuracy in 2015. Fridman also covers a genuinely surprising limitation: capsule networks expose that CNNs throw away spatial-relationship information, so two scrambled face images can look identical to a network. Good for listeners who want the ImageNet architecture lineage (AlexNet through SENet) laid out in one sitting.
Read the full episode notesMIT 6.S094: Convolutional Neural Networks for End-to-End Learning of the Driving Task
This is the lecture where CNN theory meets the road: Fridman breaks down convolutional layers, filters, and pooling, then maps them directly onto the self-driving pipeline. The standout reveal is that as of December 2016, Tesla Autopilot had driven 300 million miles with only one fatality, compared to roughly one fatality per 90 million miles for human drivers. He also flags SLAM as one of the few areas where deep learning still doesn't beat classical approaches. Recommended for anyone who wants to see CNNs applied to a real, high-stakes system rather than just benchmarks.
Read the full episode notesMIT 6.S094: Introduction to Deep Learning and Self-Driving Cars
The opening lecture of Fridman's MIT course frames the whole field around one question: is driving more like chess, which can be formally defined, or more like conversation, which requires real reasoning? Along the way he shows how an OpenAI agent playing Coast Runners exploited its reward function, circling in flames to rack up points instead of finishing the race, a clean illustration of why misaligned objectives matter. This is the right entry point if you want the conceptual scaffolding before the more technical CNN lectures later in the course.
Read the full episode notesMIT 6.S094: Deep Learning for Human Sensing
Fridman turns the CNN lens away from the road and onto the driver, using MIT's dataset of over five billion video frames, including 1.5 billion of the human face. The most counterintuitive finding: in a frustrating voice-navigation task, a smile turned out to be the strongest indicator of driver frustration, the opposite of what you'd expect from typical emotion-recognition training. Worth a listen if you're interested in CNNs applied to human behavior rather than just object detection.
Read the full episode notesMIT 6.S094: Deep Learning for Human-Centered Semi-Autonomous Vehicles
This lecture goes deeper on the driver-facing camera argument, drawing on data from 17 instrumented Teslas driving around Cambridge. Fridman repeats and expands on the frustration-versus-smiling finding from the companion lecture, and adds the wrinkle that solo drivers often don't express visible emotion at all since there's no audience to perform for, which complicates naturalistic data collection. He closes with the mystery of why deeper CNNs work better, using Conway's Game of Life as an analogy for emergent complexity. A solid pick for listeners already following the human-sensing thread across this course.
Read the full episode notesTheano Tutorial (Pascal Lamblin, MILA)
Less about CNN theory and more about how to actually build one: Lamblin, a core Theano developer at MILA, walks through the symbolic math compiler that early convolutional nets, including a LeNet architecture, were trained on. He shows how calling theano.grad returns symbolic gradient expressions for backpropagation, and how the scan operation lets you express loops for sequence models like LSTMs. This one's for the builders on the list, anyone who wants to see the actual code and computation graph behind a working CNN rather than just the concepts.
Read the full episode notesThat's eight ways into convolutional neural networks, from the inventor's own account to a full university course to the compiler code underneath it all. Browse the rest of our episode summaries to keep digging into deep learning, self-driving cars, or whatever the algorithm feeds you next.