Learn the Basics || Quick Start || Dataset & Data Structure || Learning Part || Reasoning Part || Evaluation Metrics || Bridge
Reasoning part
In this section, we will look at how to build the reasoning part, which leverages domain knowledge and performs deductive or abductive reasoning. In ABLkit, building the reasoning part involves two steps:
Build a knowledge base by creating a subclass of
KBBase, which specifies how to process pseudo-label of an example to the reasoning result.Create a reasoner by instantiating the class
Reasonerto minimize inconsistencies between the knowledge base and pseudo labels predicted by the learning part.
from ablkit.reasoning import KBBase, GroundKB, PrologKB, Reasoner
Building a knowledge base
Generally, we can create a subclass derived from KBBase to build our own
knowledge base. In addition, ABLkit also offers several predefined
subclasses of KBBase (e.g., PrologKB and GroundKB),
which we can utilize to build our knowledge base more conveniently.
Building a knowledge base from KBBase
For the user-built KB from KBBase (a derived subclass), it’s only
required to pass the pseudo_label_list parameter in the __init__ method
and override the logic_forward method:
pseudo_label_listis the list of possible pseudo-labels (also, the output of the machine learning model).logic_forwarddefines how to perform (deductive) reasoning, i.e. matching each example’s pseudo-labels to its reasoning result.
Note
Generally, the overridden method logic_forward provided by the user accepts
only one parameter, pseudo_label (pseudo-labels of an example). However, for certain
scenarios, deductive reasoning in the knowledge base may necessitate information
from the input. In these scenarios, logic_forward can also accept two parameters:
pseudo_label and x. See examples in Zoo.
After that, other operations, including how to perform abductive reasoning, will be automatically set up.
MNIST Addition example
As an example, the pseudo_label_list passed in MNIST Addition is all the
possible digits, namely, [0,1,2,...,9], and the logic_forward
should be: “Add the two pseudo-labels to get the result.”. Therefore, the
construction of the KB (add_kb) for MNIST Addition would be:
class AddKB(KBBase):
def __init__(self, pseudo_label_list=list(range(10))):
super().__init__(pseudo_label_list)
def logic_forward(self, pseudo_labels):
return sum(pseudo_labels)
add_kb = AddKB()
and (deductive) reasoning in add_kb would be:
pseudo_labels = [1, 2]
reasoning_result = add_kb.logic_forward(pseudo_labels)
print(f"Reasoning result of pseudo-labels {pseudo_labels} is {reasoning_result}.")
- Out:
Reasoning result of pseudo-labels [1, 2] is 3
Other optional parameters
We can also pass the following parameters in the __init__ method when building our
knowledge base:
max_err(float, optional), specifying the upper tolerance limit when comparing the similarity between the reasoning result of pseudo-labels and the ground truth during abductive reasoning. This is only applicable when the reasoning result is of a numerical type. This is particularly relevant for regression problems where exact matches might not be feasible. Defaults to 1e-10. See an example.use_cache(bool, optional), indicating whether to use cache to store previous candidates (pseudo-labels generated from abductive reasoning) to speed up subsequent abductive reasoning operations. Defaults to True. For more information of abductive reasoning, please refer to this.cache_size(int, optional), specifying the maximum cache size. This is only operational whenuse_cacheis set to True. Defaults to 4096.
Building a knowledge base from Prolog file
When aiming to leverage knowledge base from an external Prolog file
(which contains how to perform reasoning), we can directly create an
instance of class PrologKB. Upon instantiation of
PrologKB, we are required to pass the pseudo_label_list (same as KBBase)
and pl_file (the Prolog file) in the __init__ method.
What is a Prolog file?
A Prolog file (typically have the extension .pl) is a script or source
code file written in the Prolog language. Prolog is a logic programming language
where the logic is represented as facts
(basic assertions about some world) and
rules (logical statements that describe the relationships between facts).
A computation is initiated by running a query over these facts and rules.
See some Prolog examples
in SWISH.
After the instantiation, other operations, including how to perform abductive reasoning, will also be automatically set up.
Warning
Note that to use the default logic forward and abductive reasoning
methods in this class, the Prolog (.pl) file should contain a rule
with a strict format: logic_forward(Pseudo_labels, Res).
Otherwise, we might have to override logic_forward and
get_query_string to allow for more adaptable usage.
MNIST Addition example (cont.)
As an example, we can first write a Prolog file for the MNIST Addition
example as the following code, and then save it as add.pl.
pseudo_label(N) :- between(0, 9, N).
logic_forward([Z1, Z2], Res) :- pseudo_label(Z1), pseudo_label(Z2), Res is Z1+Z2.
Afterwards, the construction of knowledge base from Prolog file
(add_prolog_kb) would be as follows:
add_prolog_kb = PrologKB(pseudo_label_list=list(range(10)), pl_file="add.pl")
Building a knowledge base with GKB from GroundKB
We can also inherit from class GroundKB to build our own
knowledge base. In this way, the knowledge built will have a Ground KB
(GKB).
What is Ground KB?
Ground KB is a knowledge base prebuilt upon class initialization, storing all potential candidates along with their respective reasoning result. The key advantage of having a Ground KB is that it may accelerate abductive reasoning.
GroundKB is a subclass of GKBBase. Similar to KBBase, we
are required to pass the pseudo_label_list parameter in the __init__ method and
override the logic_forward method, and are allowed to pass other
optional parameters. Additionally, we are required pass the
GKB_len_list parameter in the __init__ method.
GKB_len_listis the list of possible lengths for pseudo-labels of an example.
After that, other operations, including auto-construction of GKB, and how to perform abductive reasoning, will be automatically set up.
MNIST Addition example (cont.)
As an example, the GKB_len_list for MNIST Addition should be [2],
since all pseudo-labels in the example consist of two digits. Therefore,
the construction of KB with GKB (add_ground_kb) of MNIST Addition would be
as follows. As mentioned, the difference between this and the previously
built add_kb lies only in the base class from which it is derived
and whether an extra parameter GKB_len_list is passed.
class AddGroundKB(GroundKB):
def __init__(self, pseudo_label_list=list(range(10)),
GKB_len_list=[2]):
super().__init__(pseudo_label_list, GKB_len_list)
def logic_forward(self, nums):
return sum(nums)
add_ground_kb = AddGroundKB()
Performing abductive reasoning in the knowledge base
As mentioned in What is Abductive Reasoning?, abductive reasoning enables the inference of candidates (i.e., possible pseudo-labels) as potential explanations for the reasoning result. Also, in Abductive Learning where an observation (pseudo-labels of an example predicted by the learning part) is available, we aim to let the candidate do not largely revise the previously identified pseudo-labels.
KBBase (also, GroundKB and PrologKB) implement the method
abduce_candidates(pseudo_label, y, x, max_revision_num, require_more_revision)
for performing abductive reasoning, where the parameters are:
pseudo_label, pseudo-labels of an example, usually generated by the learning part. They are to be revised by abductive reasoning.y, the ground truth of the reasoning result for the example. The returned candidates should be compatible with it.x, the corresponding input example. If the information from the inputis not required in the reasoning process, then this parameter will not have any effect.
max_revision_num, an int value specifying the upper limit on the number of revised labels for each example.require_more_revision, an int value specifying additional number of revisions permitted beyond the minimum required. (e.g., If we set it to 0, even ifmax_revision_numis set to a high value, the method will only output candidates with the minimum possible revisions.)
And it returns a list of candidates (i.e., revised pseudo-labels of the example)
that are all compatible with y.
MNIST Addition example (cont.)
As an example, with MNIST Addition, the candidates returned by
add_kb.abduce_candidates would be as follows:
|
|
|
|
Output |
|---|---|---|---|---|
[1,1] |
8 |
1 |
0 |
[[1,7], [7,1]] |
[1,1] |
8 |
1 |
1 |
[[1,7], [7,1]] |
[1,1] |
8 |
2 |
0 |
[[1,7], [7,1]] |
[1,1] |
8 |
2 |
1 |
[[1,7], [7,1], [2,6], [6,2], [3,5], [5,3], [4,4]] |
[1,1] |
11 |
1 |
0 |
[] |
As another example, if we set the max_err of AddKB to be 1
instead of the default 1e-10, the tolerance limit for consistency will
be higher, hence the candidates returned would be:
|
|
|
|
Output |
|---|---|---|---|---|
[1,1] |
8 |
1 |
0 |
[[1,7], [7,1], [1,6], [6,1], [1,8], [8,1]] |
[1,1] |
11 |
1 |
0 |
[[1,9], [9,1]] |
Creating a reasoner
After building our knowledge base, the next step is creating a reasoner. Due to the indeterminism of abductive reasoning, there could be multiple candidates compatible with the knowledge base. When this happens, reasoner can minimize inconsistencies between the knowledge base and pseudo-labels predicted by the learning part, and then return only one candidate that has the highest consistency.
We can create a reasoner simply by instantiating class
Reasoner and passing our knowledge base as a parameter. As an
example for MNIST Addition, the reasoner definition would be:
reasoner_add = Reasoner(kb_add)
When instantiating, besides the required knowledge base, we may also specify:
max_revision(int or float, optional), specifies the upper limit on the number of revisions for each example when performing abductive reasoning in the knowledge base. If float, denotes the fraction of the total length that can be revised. A value of -1 implies no restriction on the number of revisions. Defaults to -1.require_more_revision(int, optional), Specifies additional number of revisions permitted beyond the minimum required when performing abductive reasoning in the knowledge base. Defaults to 0.use_zoopt(bool, optional), indicating whether to use the ZOOpt library, which is a library for zeroth-order optimization that can be used to accelerate consistency minimization. Defaults to False.dist_func(str, optional), specifying the distance function to be used when determining consistency between your prediction and candidate returned from knowledge base. This can be either a user-defined function or one that is predefined. Valid predefined options include “hamming”, “confidence”, “avg_confidence”, “similarity” and “rejection”. For “hamming”, it directly calculates the Hamming distance between the predicted pseudo-label in the data example and candidate. For “confidence” and “avg_confidence”, it calculates the confidence distance between the predicted probabilities and each candidate, defined as1 - productand1 - averageof the candidate’s per-symbol probabilities respectively. For “similarity”, it compares candidates against the geometry of the model’s embeddings. This requires the wrapped PyTorch model to implementextract_features(x)(returning, for example, penultimate-layer activations);BasicNNthen surfaces them via its ownextract_featuresmethod, andABLModelautomatically stores the resulting embeddings ondata_example.embeddingsfor the reasoner. For “rejection”, it combines the confidence distance with a candidate-complexity penalty so that shorter candidates are favored when scores are close. Defaults to “confidence”.idx_to_label(dict, optional), a mapping from index in the base model to label.If not provided, a default order-based index to label mapping is created. Defaults to None.
The main method implemented by Reasoner is
abduce(data_example), which obtains the most consistent candidate
based on the distance function defined in dist_func.
MNIST Addition example (cont.)
As an example, consider these data examples for MNIST Addition:
# favor "1" for the first label
prob1 = [[0, 0.99, 0, 0, 0, 0, 0, 0.01, 0, 0],
[0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1]]
# favor "7" for the first label
prob2 = [[0, 0.01, 0, 0, 0, 0, 0, 0.99, 0, 0],
[0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1]]
example1 = ListData()
example1.pred_pseudo_label = [1, 1]
example1.pred_prob = prob1
example1.Y = 8
example2 = ListData()
example2.pred_pseudo_label = [1, 1]
example2.pred_prob = prob2
example2.Y = 8
The compatible candidates after abductive reasoning for both examples
would be [[1,7], [7,1]]. However, when the reasoner calls abduce
to select only one candidate based on the “confidence” distance function,
the output would differ for each example:
reasoner_add = Reasoner(kb_add, dist_func="confidence")
candidate1 = reasoner_add.abduce(example1)
candidate2 = reasoner_add.abduce(example2)
print(f"The outputs for example1 and example2 are {candidate1} and {candidate2}, respectively.")
- Out:
The outputs for example1 and example2 are [1,7] and [7,1], respectively.
Specifically, as mentioned before, “confidence” calculates the distance between the data
example and candidates based on the confidence derived from the predicted probability.
Take example1 as an example, the pred_prob in it indicates a higher
confidence that the first label should be “1” rather than “7”. Therefore, among the
candidates [1,7] and [7,1], it would be closer to [1,7] (as its first label is “1”).