Partners
Mission
At Diagonal Labs, we believe in embodied intelligence that interacts with the real world, combining high-level reasoning capabilities with on-the-field actuation.
The building blocks are here, the theory is solid, but key breakthroughs are left to be discovered.
We deeply believe in the value of this fundamental research, and the immense benefits that it promises.
model
motor ┌──────┐
command │██████│
┌────────────┤██████│
{ │ │██████│
{ } │ └──▲───┘
}_{ _│{ │
.-{ } │ }-. │
( } ┌──▼───┐) ┌──┼───┐
|`-.._│___..-│| │ ..-│
| │ ┼───────►│¯¯¯ │
| │ │__ \ │ │
| └──────┘| ) ) └──────┘
| |/ / patch
| / /
| ( /
\ y'
`-.._____..-'
Principles
The right inductive biases
We believe that key inductive biases hold the secret to intelligence in the neocortex. Our mission is to instore them in machines.
Intelligence in the neocortex is achieved through structured representations based on reference frames.
Evidence suggests that even high-level abstract reasoning and concepts are structured in reference frames.
· · · · · · · · ·
· · · · → → → · ·
· · · ↗ ↑ ↑ ↑ ↘ ·
· · ↗ ↑ ↑ ↑ ↑ ↑ ↘
· → → ↑ ↑ ↑ ↑ ↑ →
· · ↘ ↑ ↑ ↑ ↑ ↑ ↗
· · · ↘ ↑ ↑ ↑ ↗ ·
· · · · → → → · ·
· · · · · · · · ·
· · · · · · · · ·
· · ↗ ↗ ↑ ↖ ↖ · ·
· ↗ ↑ ↑ ↑ ↑ ↑ ↖ ·
↗ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↖
→ → → ↑ ↑ ↑ ← ← ←
↘ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↗
· ↘ ↓ ↓ ↓ ↓ ↓ ↗ ·
· · ↘ ↘ ↓ ↙ ↙ · ·
· · · · · · · · ·
· · · · · · · · ·
· ↘ ↓ ↓ ↓ ↓ ↓ ↙ ·
· · ↘ ↓ ↓ ↓ ↙ · ·
· · · ↘ ↓ ↙ · · ·
→ → → → ↓ ← ← ← ←
· · · ↗ ↑ ↖ · · ·
· · ↗ ↑ ↑ ↑ ↖ · ·
· ↗ ↑ ↑ ↑ ↑ ↑ ↖ ·
· · · · · · · · ·
A shift in computing paradigm
It is believed that intelligence in mammals was evolved through the replication a of semi-independent functional unit of intelligence, called a cortical column.
In other words, the neocortex is architected as a heterogeneous set of weights operating on a heterogeneous set of inputs.
This is fundamentally incompatible with the GPU parallelism paradigm. Our methods require a fundamental shift in the way we achieve scalability.
At Diagonal Labs, we draw on our expertise in computer architecture and high-performance computing to carefully ensure that our methods can fundamentally scale massively using existing computing paradigms.
An intelligent algorithm is only as good as its ability to be massively scaled on existing hardware.
patch patch patch patch
│ │ │ │
│ │ │ │
┌───▼────┐ ┌───▼────┐ ┌───▼────┐ ┌───▼────┐
│ Sensor │ │ Sensor │ │ Sensor │ │ Sensor │
│ Module ┼─┤ Module ┼─┼ Module ┼─┼ Module │
└───┬────┘ └───┬────┘ └───┬────┘ └───┬────┘
│ │ │ │
│ │ │ │
┌───▼────┐ ┌───▼────┐ ┌───▼────┐ ┌───▼────┐
│Learning│ │Learning│ │Learning│ │Learning│
│ Module ┼─┼ Module ┼─┼ Module ┼─┼ Module │
└───┬────┘ └───┬────┘ └───┬────┘ └───┬────┘
│ │ │ │
│ │ │ │
┌───▼────┐ ┌───▼────┐ ┌───▼────┐ ┌───▼────┐
│Learning│ │Learning│ │Learning│ │Learning│
│ Module ┼─┼ Module ┼─┼ Module─┼─┼ Module │
└────────┘ └────────┘ └────────┘ └────────┘
│ │ │ │
│ │ │ │
┌───▼────┐ ┌───▼────┐ ┌───▼────┐ ┌───▼────┐
│Learning│ │Learning│ │Learning│ │Learning│
│ Module ┼─┼ Module ┼─┼ Module ┼─┼ Module │
└────────┘ └────────┘ └────────┘ └────────┘
Team
We are a team with experience in AI research, neuroscience, computer architecture, high-performance computing and infrastructure.