54 Roads Not Taken
Depends on: The Transformer
54.1 Why these matter
Each idea below solved a specific, real limitation of its era’s mainline architecture and was not displaced because the problem was fake; it was displaced by a simpler alternative that scaled better.
54.2 The ideas
| Architecture | Problem targeted | Core mechanism |
|---|---|---|
| Capsule Networks (Sabour et al., 2017) | Max-pooling discards spatial pose (position, orientation, scale) | Vectors (“capsules”) encode presence and pose; routing-by-agreement replaces pooling |
| Neural Turing Machines (Graves et al., 2014) | A fixed-size RNN hidden state bottlenecks algorithmic generalization (copy, sort) | A differentiable external memory matrix with content- and location-based read/write heads |
| Neural ODEs (Chen et al., 2018) | Depth is a fixed, discrete architectural choice | A ResNet block \(y = x + F(x)\) reinterpreted as one Euler step of \(\frac{dz}{dt} = f(z(t), t)\), solved by an ODE solver |
54.2.1 Capsule Networks: Pose-Aware Routing
Capsule length encodes the probability a feature is present; orientation encodes its pose. Routing-by-agreement replaces pooling: a lower-level capsule sends its output to whichever higher-level capsule’s prediction it agrees with most, so a face capsule activates only when its eye, nose, and mouth capsules agree on a consistent pose, not merely on presence.
54.2.2 Neural Turing Machines: Differentiable External Memory
A controller network reads and writes an external memory matrix through differentiable heads, addressing memory by content (find the slot most similar to a query) or by location (move to a nearby slot). Controller and memory access train end to end with backpropagation, the same mechanism every other architecture in this book uses.
54.2.3 Neural ODEs: Continuous-Depth Networks
Stacking \(N\) discrete ResNet blocks approximates a continuous transformation via the Euler method. Replacing the discrete stack with an ODE solver makes depth an emergent property of how many solver steps a given input needs, rather than a fixed architectural choice, an adaptive, per-input compute budget.
54.2.4 What to observe
- Each targets a specific, real weakness of its era’s mainline: pooling’s spatial information loss, RNN hidden-state capacity, fixed discrete depth.
- Each introduces a genuinely new computational primitive (routing-by- agreement, differentiable memory addressing, an ODE solver) rather than recomposing existing primitives, which is precisely what made each expensive to scale.
- All three remain trainable end to end with backpropagation.
54.3 Why they did not take over
| Architecture | Why it lost |
|---|---|
| Capsule Networks | Routing-by-agreement is iterative and expensive per capsule; did not scale to the depth and data volume that made CNNs, then attention, effective through scale alone |
| Neural Turing Machines | Differentiable memory addressing was unstable to train at scale; a Transformer’s long context window is itself an attended-to memory, delivering most of the benefit without a separate module |
| Neural ODEs | Numerical ODE solvers are slower in practice than a fixed discrete stack for comparable accuracy; few tasks need adaptive per-input depth badly enough to justify the cost |