## Neural networks, Occam’s razor, and heuristic search

Neural networks sacrifice Occam’s razor in exchange for the ability to efficiently find parameters that fit the training data. This may be a fantastic trade...

Given a fixed set of proofs, it is natural to consider a lemma to be good to the extent it enables compression of the proofs. But how might we know if a lemma is good before we know how it will be used? Although I do not expect there to ever be a definitive answer to this question, a recent paper out of Chalmers University of Technology suggests a simple and perhaps even useful starting point: a lemma is good if it is minimal and non-redundant. The authors have also released a functioning Haskell library called QuickSpec that finds “good” lemmas about a set of user-provided functions. In this post, we will give a brief overview of their approach and look at a few examples.

As mentioned above, the basic premise of QuickSpec is that a lemma is good if it is minimal and non-redundant. Here *minimal* means minimal in terms of an ordering on lemma statements, and *non-redundant* means it does not follow from smaller lemmas based on first-order equational reasoning alone. QuickSpec looks only for equational lemmas, and (ignoring many crucial optimizations) it proceeds by enumerating all possible terms (which may contain free variables) in increasing order. For every term, it normalizes it with respect to the confluent rewriting system induced by the existing set of equational lemmas. If its normal form has already been visited, it is considered redundant and is discarded. Otherwise, it is evaluated on random inputs and compared to previously visited terms. If it is found to be disequal to all of them, it is simply added to the set of visited terms. On the other hand, if it “seems” to be equal to a visited term, the conjecture that the two terms are indeed equal is added to the set of minimal non-redundant (though unproven) lemmas.

For example, we can query QuickSpec on simple boolean functions as follows:

```
data Boolean = T | F deriving (Eq, Show, Ord)
instance Arbitrary Boolean where
arbitrary = do
b <- (arbitrary :: Gen Bool)
if b then return T else return F
andb T T = T
andb _ _ = F
orb F F = F
orb _ _ = T
negb F = T
negb T = F
nandb b1 b2 = negb (andb b1 b2)
main = quickSpec [
con "F" (F :: Boolean),
con "T" (T :: Boolean),
con "and" (andb :: Boolean -> Boolean -> Boolean),
con "or" (orb :: Boolean -> Boolean -> Boolean),
con "nand" (nandb :: Boolean -> Boolean -> Boolean),
con "neg" (negb :: Boolean -> Boolean),
monoType (Proxy :: Proxy Boolean)
]
```

and in under four seconds we get the following output:

```
== Functions ==
F :: Boolean
T :: Boolean
and :: Boolean -> Boolean -> Boolean
or :: Boolean -> Boolean -> Boolean
nand :: Boolean -> Boolean -> Boolean
neg :: Boolean -> Boolean
== Laws ==
1. neg F = T
2. neg T = F
3. and x y = and y x
4. and x x = x
5. nand x y = nand y x
6. nand x x = neg x
7. or x y = or y x
8. or x x = x
9. and x F = F
10. and x T = x
11. nand x F = T
12. nand x T = neg x
13. neg (neg x) = x
14. or x F = x
15. or x T = T
16. and x (neg x) = F
17. nand x (neg x) = T
18. neg (and x y) = nand x y
19. or x (neg y) = nand y (neg x)
20. and x (and y z) = and y (and x z)
21. and x (nand x y) = and x (neg y)
22. and (nand x y) (nand x z) = nand x (or y z)
```

As a second example, if we run QuickSpec with addition, subtraction and multiplication on integers, we get the following output in twenty six seconds:

```
== Functions ==
zero :: Z
one :: Z
add :: Z -> Z -> Z
neg :: Z -> Z
sub :: Z -> Z -> Z
mul :: Z -> Z -> Z
== Laws ==
1. neg zero = zero
2. add x y = add y x
3. mul x y = mul y x
4. sub x x = zero
5. add x zero = x
6. mul x one = x
7. mul x zero = zero
8. neg (neg x) = x
9. sub x zero = x
10. sub zero x = neg x
11. add x (neg y) = sub x y
12. mul x (neg y) = mul y (neg x)
13. neg (mul x y) = mul x (neg y)
14. neg (sub x y) = sub y x
15. add x (add y z) = add y (add x z)
16. mul x (add y y) = mul y (add x x)
17. mul x (mul y z) = mul y (mul x z)
18. mul x (add y one) = add x (mul x y)
19. add (mul x y) (mul x z) = mul x (add y z)
20. mul x (add y (add y y)) = mul y (add x (add x x))
21. sub (mul x x) (mul y y) = mul (add x y) (sub x y)
```

In both of the preceeding examples, I think readers will agree that these are indeed excellent lemmas.

I think QuickSpec is currently the gold standard for lemma generation, from both a practical and a philosophical perspective. Nonetheless it still has major weaknesses, most notably that it requires brute-force enumeration and so is unlikely to scale. A future post will consider the possibility of using machine learning to help speed up the search for “good” lemmas.

Neural networks sacrifice Occam’s razor in exchange for the ability to efficiently find parameters that fit the training data. This may be a fantastic trade...

Given a fixed set of proofs, it is natural to consider a lemma to be good to the extent it enables compression of the proofs. But how might we know if a lemm...

Many AI researchers (myself included) are optimistic that stochastic gradient descent (SGD) can somehow be leveraged to help learn sophisticated computer pro...

A friend of mine recently proposed the General Intelligence Hypothesis (GIH), which is a restatement of an old idea, and which has both a weak and a strong v...

Consider the following puzzle, from Martin Gardner’s My Best Mathematical and Logic Puzzles (1994):

Suppose you have a list of 25 characters, 13 of which are ‘a’, and no character ever appears twice in a row. What can you tell me about the list?

Humans can learn to become competent at playing board games such as Chess and Go, but humans are phenomenally excellent at learning how to play such games qu...

The word “abstraction” is vague in most contexts within the AI community. Has AlphaGo learned abstractions? I don’t think the question is relevant. AlphaGo ...

The AlphaZero algorithm developed by David Silver, Julian Schrittwieser, Karen Simonyan and others at DeepMind has proved very successful in the board games ...